Artificial intelligence and machine learning for simulation-based inference in B0 -> k*0 l+ l-

Abstract

We investigate neural simulation-based inference approaches to fitting the deviation of Wilson Coefficient 9 (C_9) from its Standard Model value (C_9_SM) given B0 -> K*0 mu+ mu- events simulated in the context of the Belle II experiment. We denote this deviation as delta_C_9. We compare three neural network-based approaches to this multi-dimensional fitting problem. The first approach converts the dataset into a three-dimensional grid and fits for delta_C_9 using computer vision techniques. The second approach uses the deep sets architecture to predict delta_C_9 from a dataset while enforcing the permutation invariance of events. The third approach trains a classification model to predict a binned probability distribution over delta_C_9 given a single event. Predictions are then aggregated using the independence of events to obtain the binned delta_C_9 probability distribution given the entire dataset. We train and evaluate models on simulated datasets with and without detector effects. We also train and evaluate models on a dataset that includes simulated background events from the M_bc sideband.